Statistical interpretation of machine learning-based feature importance scores for biomarker discovery

Area of Science:

• Bioinformatics
• Computational Biology
• Biostatistics

Background:

• Univariate statistical tests are common for biomarker discovery due to simplicity and interpretability.
• However, they fail to capture complex interactions between variables in biological processes.
• Machine learning offers multivariate insights but lacks statistical interpretability of relevance scores.

Purpose of the Study:

• To evaluate existing and novel procedures for extracting statistically interpretable features from machine learning relevance rankings.
• To address the challenge of determining relevance thresholds for feature selection in biomarker discovery.
• To facilitate the adoption of machine learning methods by biologists and physicians.

Main Methods:

• Evaluation of several existing and novel procedures for feature extraction from machine learning rankings.
• Transformation of machine learning relevance scores into statistically interpretable measures (e.g., p-values, FDR, FWER).
• Testing procedures on artificial datasets and real microarray data.

Main Results:

• Some evaluated procedures effectively extract truly relevant biomarkers by providing statistical interpretability.
• Methods vary in computational time and the trade-off between false positives and false negatives.
• Using model performance alone for feature selection can be counter-productive.

Conclusions:

• Statistically interpretable measures derived from machine learning relevance scores can significantly aid biomarker discovery.
• The developed methods offer practical utility for biologists and physicians in identifying informative biomarkers.
• Careful consideration of feature selection criteria is crucial, as model performance may not always be the optimal metric.